← Back to Insights
Briefing · Agent Commerce

Be Callable Before You Become Autonomous.

A verified read on what businesses must expose to software buyers, why the transaction economics remain unproved, and which authority and evidence controls must survive.

Conditional sign-off verdict

It is reasonable to expose one machine-usable purchase flow if the sponsor can name the governing authority, counterparty agreement, spend envelope, revocation path, exception owner, and transaction receipt. Do not call it a new customer channel until production evidence shows incremental completed demand or lower cycle cost after fraud, returns, disputes, integration, monitoring, approval, and exception labor.

The sign-off test

Discover

Expose authoritative product identity, price, availability, terms, seller identity, and freshness in a machine-readable form.

Authorize

Bind the principal, agent, merchant or payee, amount, goods scope, duration, approval trigger, and prohibited cases.

Transact

Preserve merchant, payment, tax, fulfillment, disclosure, and record-retention duties at the correct legal boundary.

Recover

Prove revocation, rejection, refund, dispute ownership, duplicate prevention, and receipt reconstruction separately.

Owner, briefing, proof

Owner

A business sponsor who convenes the legal, privacy, security, finance, procurement or consumer, payment, and operations authorities that apply to the use case.

Briefing

The exact purchase class, jurisdiction, counterparties, outcome, baseline, authority envelope, exception path, and two explicit no-scale conditions.

Proof

A replayable record linking principal instruction, agent identity, seller, amount, policy result, approval, order events, revocation state, and final remedy.

What leaders should take from it

1
The interface is real. The new economics are not yet proved.

OpenAI, Stripe, Visa, and Mastercard have live or pilot mechanics for discovery, checkout, agent-bound credentials, and transaction evidence. Public primary sources still do not disclose a comparable agent-originated outcome cohort.

2
Automation does not transfer commercial responsibility.

Documented flows preserve the merchant's payment, fulfillment, returns, support, communication, and merchant-of-record duties. Electronic contracting also preserves attribution, acceptance, and substantive law.

3
The delegation record is the control product.

Leading designs converge on credentials and policies bounded by a principal, agent, merchant, amount, instruction, time window, or purchase category. The strategic fight is who owns and interprets that record.

4
Start where judgment is already bounded.

Historical EDI, current sourcing automation, and consumer-choice research all point toward known counterparties, stable terms, deterministic limits, measurable baselines, and explicit exceptions before open-ended purchasing.

5
Security benchmarks are test inputs, not production loss forecasts.

Prompt-injection and product-position studies reveal real mechanisms, while their rates remain specific to models, tasks, and synthetic or benchmark environments. They do not quantify payment loss or remedy cost.

Where the evidence stops

No reviewed primary source reports agent-originated conversion, completed orders, contribution margin, protocol take rate, fraud, disputes, returns, and exception labor on one denominator. The evidence also does not establish one lawful delegation design across consumer and business purchases, jurisdictions, payment methods, or data categories.

The findings in full

9/10
Agent commerce is an interface standard, not a transfer of responsibility.

OpenAI's initial Instant Checkout let a buyer confirm an order in ChatGPT while the merchant remained merchant of record and owned payment, fulfillment, returns, support, and communication. Stripe documented a seller-side protocol lifecycle. OpenAI later shifted its checkout direction toward merchant-owned flows while retaining product discovery. The durable change is a machine interface across discovery, selection, authorization, and handoff.

9/10
A bounded, revocable delegation record is the minimum control architecture.

Visa and Mastercard center agent-bound credentials, authenticated instructions, scoped tokens, and evidence of intent. US electronic-contract and consumer-payment law preserve attribution, authority, revocation, proof, and liability. EU payment rules bind strong authentication to amount and payee where they apply.

8/10
Large transaction rails do not prove an agentic increment.

Visa and Shopify operate at enormous transaction scale, and Mastercard made a material security acquisition. Those facts identify a valuable control point. They do not measure incremental agent demand, margin, fraud, return, or exception economics. OpenAI's product shift shows the architecture remains unsettled, not that demand failed.

7/10
Bounded purchase classes are the strongest production starting point.

GAO found EDI could be very effective for routine one-to-one orders and invoices among established partners. Oracle limits autonomous sourcing through configured policy, although human approvals are not guaranteed unless enabled. Scenario experiments also found lower stated adoption intent when AI delegation reduced perceived choice or autonomy.

6/10
Attack and choice effects are measurable, with limited external validity.

InjecAgent reports benchmark-specific indirect-injection success for a GPT-4 agent. Task Shield sharply reduced targeted success in one AgentDojo condition while retaining partial utility. ACES reports product-position effects in a synthetic marketplace. None of these studies estimates real transaction loss, chargebacks, or recovery outcomes.

First moves before scaling agent commerce

01
Write the callable interface contract.

For one flow, define authoritative product, price, availability, terms, seller, checkout, payment, order-event, fulfillment, return, dispute, disclosure, consent, accessibility, retention, and receipt fields.

02
Put the authority envelope before the real approvers.

Name the principal, agent, merchant or payee, amount, goods scope, duration, approval trigger, prohibited cases, revocation path, evidence custodian, and dispute owner. A refusal keeps the flow in research or simulation.

03
Choose one contractually participating transaction class.

Use counterparties whose agreements permit electronic-agent ordering and allocate attribution and evidence, with limited spend, deterministic policy checks, and a staffed exception path.

04
Measure net economics.

Compare completed purchases or cycle time and contribution against fraud, returns, disputes, abandonment, price or inventory error, integration, monitoring, approval, and exception labor. Establish the lawful basis and data-minimization rules before collecting personal data.

05
Attack the delegation and recovery path.

In an isolated environment with written authorization and rules of engagement, test injection, manipulated ranking, stale data, amount or merchant drift, revocation, replay, duplicate orders, substitutions, partial fulfillment, returns, and disputed receipts.

06
Require two independent scale gates.

The commercial gate asks whether the flow improves the outcome after all exception cost. The assurance gate asks whether the correct authorities accept the authorization, evidence, revocation, and remedy path. One pass is not enough.

Where to start

Start with one purchase class whose baseline, approvers, contracts, counterparties, and exception owners already exist. Make that flow callable, measure its net economics, and reconstruct every approval and remedy from the receipt. Widen only when both the commercial and assurance gates pass.

Claim ledger

26/26
Checked
citation clusters traced to primary or official sources on August 25, 2026
0
Fabricated
no invented source survived verification
10
Corrected
status, scope, attribution, and legal wording narrowed
4
Demoted
conceptual, synthetic, single-company, or historical inference held out of stronger claims
ConfirmedOpenAI Instant Checkout, 2025: US Etsy launch, buyer confirmation, merchant-of-record allocation, and scoped payment-token design.openai.com
ConfirmedOpenAI product discovery, 2026: the initial checkout lacked flexibility; merchant-owned checkout and product discovery became the direction.openai.com
CorrectedStripe agentic commerce: documented ACP mechanics remain evidence, but current documentation is UCP-oriented and does not support a blanket current ACP private-preview claim.stripe.com
ConfirmedVisa Intelligent Commerce: agent-bound credentials, authenticated instructions, commerce signals, and in-development status.visa.com
CorrectedMastercard Agent Pay and Verifiable Intent: cryptographic proof of authorization comes from the later Verifiable Intent component and initial pilots, not the 2025 framework alone.mastercard.com
ConfirmedHusairi and Rossi, 2024: three randomized scenarios measured stated app-use likelihood, choice, and autonomy, not observed purchases.doi.org
CorrectedInjecAgent, 2024: 1,054 synthetic cases; GPT-4-0613 had 23.6% valid-output attack success in the base setting, not 24% of all cases.aclanthology.org
CorrectedTask Shield, ACL 2025: now peer reviewed; the AgentDojo GPT-4o results remain benchmark-specific and lack production validation.aclanthology.org
DemotedAllouah et al., 2025: ACES synthetic-market product-position effects do not establish human demand or production conversion.ace.mycustomai.io
DemotedNBER Working Paper 34468: conceptual transaction-cost analysis that explicitly leaves net welfare as an empirical question.nber.org
ConfirmedVisa FY2025: $14.2 trillion payments volume, 257.5 billion processed transactions, and $40.0 billion net revenue with definition and period limits preserved.sec.gov
ConfirmedShopify FY2025: $378.441 billion GMV, $11.556 billion GAAP revenue, and $2.007 billion non-GAAP free cash flow.sec.gov
CorrectedCFPB 2025 credit-card report: $14.3 billion in 2023 US card-fraud face-value losses. A separate lender-cost multiplier was removed from the card-fraud and merchant claim.consumerfinance.gov
DemotedMastercard Recorded Future acquisition: a $2.65 billion security investment by one network, not proof that agent commerce caused it or that all networks capture the trust layer.mastercard.com
CorrectedOracle Autonomous Sourcing Assistant: policies constrain categories, line amount, duration, and bids; approvals occur only when enabled.oracle.com
CorrectedNIST AI RMF 1.0: voluntary outcomes for human-AI roles, oversight, proficiency, and third-party risk, not a binding checklist.nist.gov
Confirmed15 U.S.C. § 7001: electronic-agent participation, attribution, acceptance rights, and electronic-record consent and retention conditions.uscode.house.gov
CorrectedRegulation E and EFTA: consumer-account EFT authority and liability; EFTA § 909(b), not the cited regulation sections alone, supplies the financial institution's proof burden.consumerfinance.gov
ConfirmedEU Delegated Regulation 2018/389: dynamic linking to amount and payee where strong customer authentication applies.eur-lex.europa.eu
ConfirmedCJEU C-249/21: the online order button itself must unambiguously indicate the payment obligation for covered consumer distance contracts.eur-lex.europa.eu
CorrectedDirective EU 2024/2853: defective-product liability for software including AI after December 8, 2026, with implementation and scope limits. It is not a general AI-liability regime.eur-lex.europa.eu
Confirmed18 U.S.C. § 1030: adversarial testing is therefore confined to isolated environments under written authorization and rules of engagement.uscode.house.gov
ConfirmedGDPR: lawful basis, data minimization, retention, and impact-assessment triggers condition any qualifying EU production-cohort measurement.eur-lex.europa.eu
CorrectedGAO NSIAD-97-26: EDI can be very effective for routine one-to-one orders and invoices; FACNET's one-to-many requirements materially differed.gao.gov
ConfirmedGAO Decision B-245714: properly secured EDI may create recordable federal obligations when it supplies assurance comparable to paper formation.gao.gov
DemotedDOJ anti-steering litigation, 2010: $1.6609 trillion in network purchase volume and about $35 billion in annual US merchant credit-card acceptance cost. The trust-layer tollbooth interpretation is historical inference.justice.gov
Contested signals: held out of the verdict

The evidence does not show that agent commerce already produces material incremental GMV, conversion, or margin; that any benchmark supplies a production fraud or loss rate; or that one delegation design transfers across transaction types and jurisdictions. Protocol activity is an adoption signal, not a business outcome.

What would change this conclusion

Related work

Verified research · 26 citation clusters checked · 0 fabricated · 10 corrected · 4 demoted